LOAD SEDIMENTS QUANTIFICATION IN ALGERIAN NORTH- WEST BASINS BY ARTIFICIAL NEURONS NETWORK METHOD

Yacine HAFIED , Nadir MAROUF , Mohamed Tewfik BOUZIANE , Boualem REMINI , Shiynkaruk LUBOMIR 4 Laboratory of Hydraulic, Biskra University, Biskra 7000, Algeria, Laboratory of Hydraulic, Oum El Bouaghi University, Algeria Department of Water Science, Faculty of Technology, Blida 1 University, Blida 9000, Algeria, National University of Water Supply and Environmental Management, Ukraine E-mail:hafied.yacine@yahoo.fr


INTRODUCTION
Water resources development in the north of Algeria requires water reservoirs construction for various purposes; however, the design and successful exploitation of water reservoirs can only be made through an accurate and proper sedimentary volume prediction. This volume estimation will have a big economic impact, knowing that an underestimation or overestimation decreases or increases the dead-volume. Underestimation can affect dam's life time by contrast, overestimation can lead to a huge increase in dam's cost.
This study is based on historical data "liquid flow rates-solid flow rates" of the North-West Algerian basins namely Cheliff, Tafna, Macta, and Oran, in order to provide the most appropriate global sedimentary load estimation model.

PRESENTATION OF THE STUDY AREAS
The study focuses on the basins located in the north-west of Algeria, i.e. Cheliff, Tafna, Macta, and Oran basins ( Figure 1). Seven (07) measurement stations were selected in this area, on the large amount of data stored, to have reliable and concrete models for the suspended sediment load quantification. The choice of these hydrometric stations is aroused by the availability and richness of historic measurement data, in our case, liquid flow rates, solid flow rates. The data were provided by National Agency of Water Resources 'Agence Nationale des Ressources Hydriques' (ANRH) in 2016 [32] -see Table 1.

BASINS CHARACTERISTICS
The characteristics of the studied watersheds are given in Table 2.

CHELIFF BASIN
It is located in the North-West region of Algeria, bordered at North by Algiers coastal basin; South by Sahara basin, West by Oran costal basin, Macta and Oran highlands and East by Chott-Melghir, Zahrez, Chott-Hodna and Isser. The main watercourse in this Basin is Chéliff-stream water.


Cheliff stream water is a river of 725 km long, located in the north-west of Algeria which is fed from the Tellian Atlas, more precisely in the Djebel AMOUR and flows into the Mediterranean Sea. The Cheliff is characterized by a very fertile valley and it is the most important river in Algeria ( Figure 3).

TAFNA BASIN
The Tafna Basin is located in the extreme north-west of Algeria while a part of it is immersed in Morocco, bordered to the north by the Oran Basin Coastal, east by Macta Basin, and west by Moroccan borders. Its main watercourse is the Tafna River.

MACTA BASIN
It is located in the north-western region of Algeria, bordered to the north by the Oran's coastal basin. One part of it extends to the Mediterranean Sea; south is bounded by the Oran's Highlands Basin; east by the Cheliff basin and west by the Tafna Basin. It is drained by two main watercourses, the Mebtouh stream water to the west and El-Hammam stream water to the east.

Distribution of Solid flow rate depending on liquid flow rate.
The representation of the measured solid flow rates in relation with the measured liquid flow rate of each studied hydrometric station is illustrated in the graphs mentioned in Figure 9.

Application of ANN method
The ANN method is a harmonized calculation model whose design is very schematically inspired from the functioning of human biological neurons. The formal neuron is designed as a PLC (programmable logic controller) with a transfer function that transforms its input to output according to specific rules. Neurons are also associated in networks whose topology of connections is variable; the efficiency of signal transmission from one neuron to another can vary: we speak about "synaptic weight", and these weights can be modulated by training rules (which presents the synaptic plasticity of biological networks). There are several methods of ANN calculations. We opted for the method of 'feedforward backprop' because it is the most appropriate [30].
An elementary neuron with R inputs is shown in Diagram 1. Each entry is weighted with w. The sum of weighted inputs and bias (b) is the input to the transfer function (f). Neurons can use any differentiable transfer function f to generate their output.

Diagram 1. Elementary Neuron
For multi-layer networks, the number of layers determines the index on the weight matrix. This network can be used as a general function approximator. It can approximate any function with a finite number of discontinuities arbitrarily well, if enough neurons are given in the hidden layer [30].

Preparation of liquid-solid flowrate (Ql-Qs) data for multilayer neural networks
Before starting the network design process, we must first collect and prepare data (Ql-Qs). It is important that the data (Ql-Qs) cover the range of inputs for which the network will be used. After collecting the data, there are two steps that must be performed before using the data (Ql-Qs) for network learning: the data shall be preprocessed and divided into sub-sets.

Neural Network Input-Output functions
Training the neural network can be more efficient if we perform certain pre-treatment steps on the inputs and targets of the network. The normalization step is applied to both the input and the target vectors in the data set (Q1-Qs). Hence, the network output always falls within a normalized range. The network output can then be transformed back into the original target data units when the network is put in use in the field.

(Ql-Qs) Data Division for optimal training on neural networks
The general practice is to divide the data into three subsets. The first subset is the training set which is used to calculate the gradient and update the weights (w) and the network bends. The second subset is the validation set. The error on the validation set is monitored during the training process. Normally, the validation error decreases during the initial phase of training, as does the error of the training set. However, when the network begins to overload the data, the error on the validation set usually begins to increase. Weights and network biases are recorded as a minimum of the validation error. The test error is not used during training, but it is used to compare different models. The chosen ratios for training, testing and validation are 0.6, 0.2 and 0.2, respectively.

Running and application of a multilayer neuron network
The fastest training function is usually the trainlm. We are now launching our ANN network for each measurement station with Ql as input and Qs as target, based on the calculation steps mentioned earlier.

Improvement of results
If the formed network is not precise enough (does not give accurate results), we try to reset the network and run it again. Each time a feedforward network is initialized; the network settings are different and can produce different solutions. As a second approach, one can increase the number of neurons hidden above 10. A larger number of neurons in the hidden layer provides more flexibility to the network. A third option is to try different training functions. Table 3 shows the calculation parameters that we have achieved after several attempts, in order to have better results.

ANN method results
We describe the variation of simulated solid flow rate simulated (Qssim) by ANN model depending on solid flowrate that was measured in each hydrometric station. The attached graphs illustrate the (Qssim-Qs) variation obtained at each measuring station ( Figure 10).   The obtained results are shown in Figure 11. As shown in the graphs obtained by the global ANN model, we can observe quasi-linear fit in the three execution blocks (training, validation and test). The correlation coefficient approaches to 1 and has the lowest correlation of about 86.6% in the test graph. Therefore, it can be deduced that the general model performed is reliable and prevailing. The representations of solid flowrate simulated of global model (Qssimglob) depending on each individual model show a good linear trend. The correlation exceeds 77% in all the measuring stations. It reaches a maximum correlation of (97%) at the Turgo Nord station in the Oran's basin ( Figure 12) and a minimum trend of (77%) at the Pierre du Chat station in the Tafna basin. These adequate results obtained approve the general simulation model found and confirm its reliability of estimation in all the hydrometric measurement stations despite the training sample did not exceed 70%.

CONCLUSION
The quantification of the suspended sediment load by the ANN method has been applied in the North-West Algerian basins including Cheliff, Tafna, Macta and Oran's. This method has proved its computing efficiency. The obtained results are almost perfect and show ideal correlations. This model has been tested on several historical data of liquid flowrates of each hydrometric station to guarantee its reliability over time. The extracted global model is valid for estimating suspended solid transport in all studied watersheds. The other established individual models are designed for each studied basin and specific to each measuring station. So, this approach allowed calculation of the solid flowrate in a concrete, precise or even eloquent way. Consequently, it makes it easier for engineers and managers of hydraulic structures to have better estimate solid transport volumes and to predict their protection. In the same way, it also allowed to gain time and means by properly estimating the sediment load and minimizing the project cost. The final global model only needs liquid flowrates as data at the program input to predict the amount of sediment transported in stream waters. We can also make measurements on site directly using the introduction of software (global model) in hardware (flow meter) something that allows reading the solid flowrate directly on the meter.